Descriptive Analysis & Preliminary Results
Introduction
The following document summarises the progress made thus far on Chapter 1: Local Fiscal Risks of Decarbonisation of my DPhil. The work aims to pursue a better understanding of how industrial transformation impacts local well-being. From an original interest in looking at all aspects of local public finance, the project has narrowed to focus on expenditure on public education and its connection to industrial prosperity and transformation.
Below, I provide some descriptive statistics and figures about the data as well as some preliminary efforts to explore the trends in these values.
The main research question is: How does industrial transformation/activity impact county-level expenditure on public education?
CZ-level Education Expenditure
Overall, expenditure on education per pupil has increased slightly in real USD.
The following plot displays each county’s time series of expenditure (total and per pupil log values).
Below are scatterplots and diagrams depicting key relationships between dependent and control variables as well as shares/components of key variables.
Most important to tease out before modelling is how different revenue sources (federal, state, county (own), and other local sources) interplay. County-level revenue for public education is a combination of both local and intergovernmental sources. The local portion of the share is almost entirely sourced through property taxes. The intergovernmental sources come from state, federal, and local aid.
Below chart plots CZ-level mean (taken over the panel time horizon) of different intergovernmental (IG) sources versus own-source revenue (generated from local sources). The solid black line represents a best-fit line and the dashed line represents a 45 degree line. The blue plot shows Total IG Revenue (Federal + State + Other Local) versus own revenue. There is a near-1:1 negative correlation between the two (ie. they are near-substitutes for one another). This effect is dominated by State IG revenue (as can be seen in the purple panel). Propoerty taxes have a near 1:1 relationship with own revenue confirming that property taxes make up own revenue sources almost completely.
The following plots show the share of revenue for public education that comes from own sources, local intergovernmental, state intergovernmental, and federal intergovernmental by state. Almost all states have a near 50% split of revenue from state and own sources, which aligns with data from the Congressional Research Service cited in the Transfer of Status report. Massachusetts has an unusually high share of local IG support (inter-school aid), eclipsing own sources almost completely. From further research, I believe this has to do with a unique structure of Massachusetts public school funding which is reliant on several multi-county funding agencies (similar to the mentioned ESAs). This anomaly might warrant the exclusion of MA from the analysis.
The below plot provides the same information as above but on a national level.
Conclusion: Corroborates the near-even split of revenue between state and own sources which seems to be a fact of public education revenue. I mainly include this as an alternative summarising figure to the above.
We know the majority of “own source” revenue comes from property taxes. The below scatterplots demonstrate the relationship between education revenue and expeniture versus property taxes collected. The solid black line represents a best-fit line and the dashed line represents a 45 degree line (intercept-adjusted to match data). The vertical dotted line represents a potential preliminary cutoff point for outliers - will be subject to further more rigorous testing.
The solid black line represents a best-fit line and the dashed line represents a 45 degree line.
GDP and Property taxes have a positive linear correlation, to be expected.
The solid black line represents a best-fit line and the dashed line represents a 45 degree line (intercept-adjusted to match data).
GDP and Education expenditure have a positive linear correlation, as expected.
If we use a Chmura economic diversity indicator and plot GDP, private industry GDP, education expenditure, and elementary education expenditure we see that there is little difference in the mean of expenditure, however higher values of each occur only in more diverse counties. This data is only available at county level, not commuting zone level.
Around 2007, many states instituted Educational Service Agencies (ESAs) which sought to “equalise” public education across the country. To date, ~45 states have ESAs which are responsible for multiple school districts, most often across multiple counties. Therefore, when modelling county-level expenditure it will be important to understand how this change in educational expenditure affected county-level spending (ie. did ESAs replace or supplement county-level funding).
Only 593 counties of 2740 in the dataset have recorded revenue/expenditure from ESAs. After some digging, I believe that these values for ESAs are improperly recorded in the sense that the revenue is recorded only in counties in which the ESA’s headquarters is located and not partitioned to the counties to which the revenue ultimately flows. All county-level total expenditure/revenue values that are used in the regressions on this page have explicitly excluded recorded ESA values for this reason (they are instead recorded in a variable called esa_tot_exp or esa_tot_rev in total and per pupil values).
The below graphs show some relationships between ESA and county-level finances. I have yet to arrive at a definitive understanding of how ESAs and county-level finances interact. They do not appear to be substitutes. As is evident, the data on ESA expenditure is patchy and highly volatile by county. At the moment, I believe this is because of imperfect/inconsistent reporting in comparison to traditional school district reporting. The four states that have recorded ESA expenditure before 2007 are California, Illinois, Minnesota, Oregon.
With this we want to demonstrate how ESAs interact with our expenditure indicator. Does high ESA spending imply low/high local spending? It seems from the below that the two have a no correlation, implying little substitution? (Each point is a county in a particular year starting from 2007, colour represents state). Warm color-scale is (top-left panel) is expenditure values whereas the cool color scales are revenue values. The black line is a 45 degree line.
In this sanity check, I look at what proportion of reported school age children per state seem to be represented in the enrollment numbers in our Survey of Local Finances dataset. We see that this share varies in magnitude by state (Maine having only 40% of school-age children covered). Considering that ~ 9% of students are enrolled in private schools across the US, the coverage in the top states is not too bad!
Trends
What would be great is to be able to econometrically test when a county is “declining.” In the first step, it would be good to identify when a county is declining overall (GDP, poverty, etc) but ideally eventually apply this to the education outcome. My hope is that being able to identify counties that are “declining” we can either use this variable as a covariate or as a central point of analysis. The below analysis looks at state- and commuting zone level variables.
Some current options:
- Deviation from t_0
- Change in logs
- y ~ beta*t
- Coefficient Fixed Effect Step Indicator Saturation (Breaks in the coefficient for each individual)
Exhibit national or state-level GDP growth as global common factor; counties with negative factor loadings might arguably exhibit decline (_)
Jennie’s suggestion:
- plot county- , CZ, or state-level deviations from a national trend
One way to look at this descriptively is to look at the national, state and county-level trends. Here the “trend” is simply the difference between expenditure per pupil at the relevant level and expenditure per pupil in the first available observation as represented here: https://www.princeton.edu/~erossi/DTNLC.pdf. (Note: this measure is imperfect for many reasons).
Trend in log Elementary Education Expenditure (growth rate). The below plots show the mean of the change in log of the variable of interest. In the first panel I include no additional controls but in the latter two panels I include the state and national level growth rates as controls.
Trend in log Elementary Education Expenditure. The below plots show the coefficient estimate on the time-step variable in the trend equation. In the first panel I include no additional controls but in the latter two panels I include the state and national level growth rates as controls.
Mean change in log Real GDP (growth rate). The below plots show the coefficient estimate on the time-step variable in the trend equation. In the first panel I include no additional controls but in the latter two panels I include the state and national level growth rates as controls.
The following isoltes those counties that have significant negative or positive trend values. ::: panel-tabset #### Negative
Positive
:::
Supplementary Materials
- National data on GDP, UER, (and poverty)
State data on GDP, UER, and poverty
Extract trends of state and national data
Deviations at county level from national and state-level trends
Plot deviations
Saturation to find structural breaks in these deviations
Result: UE, GDP, and Poverty gaps
PCA on those three gaps
Retain index from PCA
Step break in index
Line up with any notable events?
Place “gaps” in a panel structure for break detection
Autocorrelation plot to test no residual autocorrelation in model above with PCA indices.
Cointegration testing
Suggestions from Jennie:
Engle-Granger Test - bivariate scenario
Johanson Test - multivariate setting [GDP, Property Taxes, Ed. Spending]
Autocorrelation
First-differences
Autocorrelation (Education only) (X)
Correlation plots (X)
Cross-correlation plot of Education and GDP (X)
Are GDP and Education Expenditure cointegrating (or do they exhibit a decoupling?) (_)
Autocorrelation
First-differences
Correlation Plots (Education and GDP)
Cross-correlation of Educ Exp and Priv. Industry GDP
Autocorrelations of series 'X', by lag
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0
-0.051 -0.019 -0.225 -0.376 -0.146 -0.043 0.316 0.632 0.555 0.293 0.001
1 2 3 4 5 6 7 8 9 10
-0.418 -0.199 -0.196 -0.028 0.230 -0.099 -0.118 -0.016 -0.081 -0.040
Autocorrelations of series 'X', by lag
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0
-0.224 -0.172 -0.120 0.009 0.227 0.532 0.439 0.313 -0.038 -0.064 0.087
1 2 3 4 5 6 7 8 9 10
0.013 -0.068 -0.104 -0.184 -0.154 -0.038 -0.120 0.002 -0.327 -0.170
Autocorrelations of series 'X', by lag
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0
-0.373 -0.107 -0.201 -0.145 -0.145 -0.191 0.203 0.343 0.581 0.698 0.237
1 2 3 4 5 6 7 8 9 10
-0.147 -0.061 -0.198 -0.115 -0.235 -0.285 -0.131 -0.256 -0.084 -0.002
Autocorrelations of series 'X', by lag
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0
-0.004 -0.036 -0.164 -0.287 -0.272 -0.127 0.036 0.141 0.365 0.557 0.101
1 2 3 4 5 6 7 8 9 10
0.115 0.171 0.051 -0.189 -0.155 -0.091 -0.201 -0.250 -0.061 0.030
Autocorrelations of series 'X', by lag
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0
0.012 0.012 -0.046 0.041 -0.176 0.032 0.028 0.345 0.239 0.436 0.116
1 2 3 4 5 6 7 8 9 10
0.000 -0.203 -0.051 -0.068 0.023 -0.325 -0.138 -0.115 -0.067 -0.216
Autocorrelations of series 'X', by lag
-10 -9 -8 -7 -6 -5 -4 -3 -2 -1 0
0.085 -0.158 0.351 -0.392 -0.041 0.600 -0.462 -0.188 0.156 0.079 -0.018
1 2 3 4 5 6 7 8 9 10
-0.282 0.231 0.100 -0.187 0.053 0.052 0.095 0.014 0.128 -0.062